Merge branch 'main' of github.com:pytorch/torchcodec into nvdecclean

Author: NicolasHug

.github/workflows/docs.yaml

@@ -51,10 +51,10 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-          # 3.9 corresponds to the minimum python version for which we build
+          # 3.10 corresponds to the minimum python version for which we build
           # the wheel unless the label cliflow/binaries/all is present in the
           # PR.
-        python-version: ['3.9']
+        python-version: ['3.10']
         cuda-version: ['12.6']
         ffmpeg-version-for-tests: ['7']
     container:

.github/workflows/linux_cuda_wheel.yaml

@@ -61,12 +61,12 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-          # 3.9 corresponds to the minimum python version for which we build
+          # 3.10 corresponds to the minimum python version for which we build
           # the wheel unless the label cliflow/binaries/all is present in the
           # PR.
           # For the actual release we should add that label and change this to
           # include more python versions.
-        python-version: ['3.9']
+        python-version: ['3.10']
         # We test against 12.6 and 12.9 to avoid having too big of a CI matrix,
         # but for releases we should add 12.8.
         cuda-version: ['12.6', '12.9']

.github/workflows/linux_wheel.yaml

@@ -62,7 +62,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.9']
+        python-version: ['3.10']
         ffmpeg-version-for-tests: ['4.4.2', '5.1.2', '6.1.1', '7.0.1']
     needs: build
     steps:

.github/workflows/macos_wheel.yaml

@@ -63,7 +63,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.9']
+        python-version: ['3.10']
         ffmpeg-version-for-tests: ['4.4.2', '5.1.2', '6.1.1', '7.0.1']
     needs: build
     steps:

.github/workflows/reference_resources.yaml

@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
   pull_request:
     paths:
-      - test/generate_reference_resources.sh
+      - test/generate_reference_resources.py
       - .github/workflows/reference_resources.yaml # self reference
   schedule:
     - cron: '0 0 * * 0'  # on sunday
@@ -19,7 +19,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.9']
+        python-version: ['3.10']
         ffmpeg-version-for-tests: ['4.4.2', '5.1.2', '6.1.1', '7.0.1']
     steps:
       - name: Setup conda env
@@ -38,7 +38,7 @@ jobs:
       - name: Update pip
         run: python -m pip install --upgrade pip
 
-      - name: Instal generation dependencies
+      - name: Install generation dependencies
         run: |
           # Note that we're installing stable - this is for running a script where we're a normal PyTorch
           # user, not for building TorhCodec.
@@ -50,4 +50,4 @@ jobs:
 
       - name: Run generation reference resources
         run: |
-          test/generate_reference_resources.sh
+          python test/generate_reference_resources.py

.github/workflows/windows_wheel.yaml

@@ -36,19 +36,20 @@ jobs:
       with-rocm: disable
       with-cuda: disable
       build-python-only: "disable"
+      python-versions: '["3.10", "3.11", "3.12", "3.13"]'
+
 
   build:
     needs: generate-matrix
     strategy:
       fail-fast: false
     name: Build and Upload Windows wheel
-    # TODO: use @main
-    uses: nicolashug/test-infra/.github/workflows/build_wheels_windows.yml@build-platform-windows
+    uses: pytorch/test-infra/.github/workflows/build_wheels_windows.yml@main
     with:
       repository: pytorch/torchcodec
       ref: ""
-      test-infra-repository: nicolashug/test-infra
-      test-infra-ref: build-platform-windows
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
       pre-script: packaging/pre_build_script.sh
       # post-script: packaging/post_build_script.sh  TODO: consider enabling post-build checks for Windows
@@ -60,3 +61,74 @@ jobs:
       # The BUILD_AGAINST_ALL_FFMPEG_FROM_S3 var, needed to build the wheel, is
       # set in vc_env_helper.bat Couldn't find a way to set it from here.
       build-command: "python -m build --wheel -vvv --no-isolation"
+
+  install-and-test:
+    runs-on: windows-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ['3.10']
+        # TODO: FFmpeg 5 on Windows segfaults in avcodec_open2() when passing
+        # bad parameters.
+        # See https://github.com/pytorch/torchcodec/pull/806
+        ffmpeg-version-for-tests: ['4.4.2', '6.1.1', '7.0.1']
+    needs: build
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: pytorch_torchcodec__${{ matrix.python-version }}_cpu_x64
+          path: pytorch/torchcodec/dist/
+      - name: Setup conda env
+        uses: conda-incubator/setup-miniconda@v2
+        with:
+          auto-update-conda: true
+          miniconda-version: "latest"
+          activate-environment: test
+          python-version: ${{ matrix.python-version }}
+      - name: Update pip
+        run: python -m pip install --upgrade pip
+      - name: Install PyTorch
+        run: |
+          python -m pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+      - name: Install torchcodec from the wheel
+        run: |
+          wheel_path=`find pytorch/torchcodec/dist -type f -name "*.whl"`
+          echo Installing $wheel_path
+          python -m pip install $wheel_path -vvv
+      - name: Check out repo
+        uses: actions/checkout@v3
+      - name: Install ffmpeg, post build
+        run: |
+          # Ideally we would have checked for that before installing the wheel,
+          # but we need to checkout the repo to access this file, and we don't
+          # want to checkout the repo before installing the wheel to avoid any
+          # side-effect. It's OK.
+          source packaging/helpers.sh
+          assert_ffmpeg_not_installed
+          conda install "ffmpeg=${{ matrix.ffmpeg-version-for-tests }}" -c conda-forge
+          ffmpeg -version
+      - name: Test torchcodec import after FFmpeg installation
+        run: |
+          echo "Testing torchcodec import after FFmpeg is installed and PATH is updated..."
+          python -c "import torchcodec; print('TorchCodec import successful!')"
+      - name: Install test dependencies
+        run: |
+          # Ideally we would find a way to get those dependencies from pyproject.toml
+          python -m pip install numpy pytest pillow
+      - name: Delete the src/ folder just for fun
+        run: |
+          # The only reason we checked-out the repo is to get access to the
+          # tests. We don't care about the rest. Out of precaution, we delete
+          # the src/ folder to be extra sure that we're running the code from
+          # the installed wheel rather than from the source.
+          # This is just to be extra cautious and very overkill because a)
+          # there's no way the `torchcodec` package from src/ can be found from
+          # the PythonPath: the main point of `src/` is precisely to protect
+          # against that and b) if we ever were to execute code from
+          # `src/torchcodec`, it would fail loudly because the built .so files
+          # aren't present there.
+          rm -r src/
+          ls
+      - name: Run Python tests
+        run: |
+          pytest test -vvv

README.md

@@ -130,7 +130,7 @@ The following table indicates the compatibility between versions of
 
 | `torchcodec`       | `torch`            | Python              |
 | ------------------ | ------------------ | ------------------- |
-| `main` / `nightly` | `main` / `nightly` | `>=3.9`, `<=3.13`   |
+| `main` / `nightly` | `main` / `nightly` | `>=3.10`, `<=3.13`   |
 | `0.6`              | `2.8`              | `>=3.9`, `<=3.13`   |
 | `0.5`              | `2.7`              | `>=3.9`, `<=3.13`   |
 | `0.4`              | `2.7`              | `>=3.9`, `<=3.13`   |

benchmarks/decoders/benchmark_decoders_library.py

@@ -47,6 +47,9 @@ def decode_first_n_frames(self, video_file, n):
     def decode_first_n_frames_description(self, n) -> str:
         return f"first {n} frames"
 
+    def init_decode_and_resize(self):
+        pass
+
     @abc.abstractmethod
     def decode_and_resize(self, video_file, pts_list, height, width, device):
         pass
@@ -107,9 +110,12 @@ def __init__(self, backend):
         self._backend = backend
         self._print_each_iteration_time = False
         import torchvision  # noqa: F401
-        from torchvision.transforms import v2 as transforms_v2
 
         self.torchvision = torchvision
+
+    def init_decode_and_resize(self):
+        from torchvision.transforms import v2 as transforms_v2
+
         self.transforms_v2 = transforms_v2
 
     def decode_frames(self, video_file, pts_list):
@@ -267,6 +273,7 @@ def __init__(
         self._color_conversion_library = color_conversion_library
         self._device = device
 
+    def init_decode_and_resize(self):
         from torchvision.transforms import v2 as transforms_v2
 
         self.transforms_v2 = transforms_v2
@@ -379,6 +386,7 @@ def __init__(
         self._seek_mode = seek_mode
         self._stream_index = int(stream_index) if stream_index else None
 
+    def init_decode_and_resize(self):
         from torchvision.transforms import v2 as transforms_v2
 
         self.transforms_v2 = transforms_v2
@@ -443,6 +451,7 @@ def __init__(
         self._device = device
         self._seek_mode = seek_mode
 
+    def init_decode_and_resize(self):
         from torchvision.transforms import v2 as transforms_v2
 
         self.transforms_v2 = transforms_v2
@@ -546,10 +555,12 @@ def __init__(self, stream_index: str | None = None):
 
         self.torchaudio = torchaudio
 
+        self._stream_index = int(stream_index) if stream_index else None
+
+    def init_decode_and_resize(self):
         from torchvision.transforms import v2 as transforms_v2
 
         self.transforms_v2 = transforms_v2
-        self._stream_index = int(stream_index) if stream_index else None
 
     def decode_frames(self, video_file, pts_list):
         stream_reader = self.torchaudio.io.StreamReader(src=video_file)
@@ -883,6 +894,7 @@ def run_benchmarks(
 
             if dataloader_parameters:
                 bp = dataloader_parameters.batch_parameters
+                decoder.init_decode_and_resize()
                 description = (
                     f"concurrency {bp.num_threads}"
                     f"batch {bp.batch_size}"

examples/decoding/file_like.py

@@ -215,7 +215,7 @@ def read(self, size: int) -> bytes:
         self.num_reads += 1
         return self._file.read(size)
 
-    def seek(self, offset: int, whence: int) -> bytes:
+    def seek(self, offset: int, whence: int) -> int:
         self.num_seeks += 1
         return self._file.seek(offset, whence)
 

setup.py

@@ -135,9 +135,14 @@ def _build_all_extensions_with_cmake(self):
             ["cmake", str(_ROOT_DIR)] + cmake_args, cwd=self.build_temp
         )
         print("Calling cmake --build", flush=True)
-        subprocess.check_call(["cmake", "--build", "."], cwd=self.build_temp)
+        subprocess.check_call(
+            ["cmake", "--build", ".", "--config", cmake_build_type], cwd=self.build_temp
+        )
         print("Calling cmake --install", flush=True)
-        subprocess.check_call(["cmake", "--install", "."], cwd=self.build_temp)
+        subprocess.check_call(
+            ["cmake", "--install", ".", "--config", cmake_build_type],
+            cwd=self.build_temp,
+        )
 
     def copy_extensions_to_source(self):
         """Copy built extensions from temporary folder back into source tree.
@@ -156,7 +161,7 @@ def copy_extensions_to_source(self):
             #   https://stackoverflow.com/a/2339910
             extensions = ["dylib", "so"]
         elif sys.platform in ("win32", "cygwin"):
-            extensions = ["dll"]
+            extensions = ["dll", "pyd"]
         else:
             raise NotImplementedError(f"Platform {sys.platform} is not supported")